RE: [RFC PATCH] ethdev: advertise flow restore in mbuf

[Ori Kam <orika@nvidia.com>]

Hi David,

> -----Original Message-----
> From: David Marchand <david.marchand@redhat.com>
> Sent: Thursday, June 1, 2023 12:43 PM
> Subject: Re: [RFC PATCH] ethdev: advertise flow restore in mbuf
> 
> On Thu, Jun 1, 2023 at 11:31 AM Ori Kam <orika@nvidia.com> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: David Marchand <david.marchand@redhat.com>
> > > Sent: Thursday, June 1, 2023 11:48 AM
> > >
> > > On Wed, May 24, 2023 at 8:44 PM David Marchand
> > > <david.marchand@redhat.com> wrote:
> > > > On Wed, May 24, 2023 at 6:00 PM Ori Kam <orika@nvidia.com> wrote:
> > > > > > As reported by Ilya [1], unconditionally calling
> > > > > > rte_flow_get_restore_info() impacts an application performance for
> > > drivers
> > > > > > that do not provide this ops.
> > > > > > It could also impact processing of packets that require no call to
> > > > > > rte_flow_get_restore_info() at all.
> > > > > >
> > > > > > Advertise in mbuf (via a dynamic flag) whether the driver has more
> > > > > > metadata to provide via rte_flow_get_restore_info().
> > > > > > The application can then call it only when required.
> > > > > >
> > > > > > Link: http://inbox.dpdk.org/dev/5248c2ca-f2a6-3fb0-38b8-
> > > > > > 7f659bfa40de@ovn.org/
> > > > > > Signed-off-by: David Marchand <david.marchand@redhat.com>
> > > > > > ---
> > > > > > Note: I did not test this RFC patch yet but I hope we can resume
> and
> > > > > > maybe conclude on the discussion for the tunnel offloading API.
> > > > > >
> > > > >
> > > > > I think your approach has a good base but what happens if
> > > > > it is not relevant for all flows? In this case your solution will not work.
> > > >
> > > > Sorry, I am not following.
> > > > Could you develop?
> > >
> > > I still don't get your comment, could you give an example/usecase
> > > where this approach can't work?
> > > Thanks.
> > >
> > I'm think of a use case that some flows have the restore info, while
> > other don't for example, we get arp packets or some packets that
> > are not tunneled, and we also get tunneled packets.
> >
> > Or for example PMD supports this flag but the application didn't offload
> such a rule yet.
> 
> Again, maybe I missed something, but my proposal is for a *per packet*
> report from the driver.
> I am not for a global driver capability, if this is what you have in mind.
> 
My bad, per packet solves the issue I was talking about, but it makes it worse.
This means that PMD needs to add logic in it's datapath. This may affect all
traffic. 

> 
> >
> > In those cases application will be slow even if he didn't offload the rules,
> > I assume we can say that if application wants to use this he should know
> > that other packets will have some performance issues.
> >
> > From my point of view if application requested the tunnel offload it should
> > always check this function.
> 
> With a per packet flag, the application only calls restore_info when
> such tunnel offload rules have been requested, and only for packets
> that require it.
> 
> 
> --
> David Marchand